Is there a pre-trained model？

[wchuan163]

Is there any pre-trained model？

[johndpope]

the closest to anything is recreation of the foundation paper that vasa is built on - MegaPortraits https://github.com/johndpope/MegaPortrait-hack/issues/36 Before throwing more time at this MegaPortrait - EmoPortrait code / model is supposed to drop. This guy @neeek2303 (who now works at Meta) was key author of the original samsung paper / so he has definitive code / correct model structure https://github.com/neeek2303/EMOPortraits

kinda of in a holding pattern for this to drop - then circle back to do diffusion stuff. I drafted this a while back - https://github.com/johndpope/VASA-1-hack/pulls

more recently been playing around with another microsoft paper https://github.com/johndpope/IMF it's stylegan2 based - learned a lot reimplementing this one.

this week i switch back to stable diffusion.

Also - finally have some compute on google vertex https://github.com/johndpope/vertex-jumpstart

[mantrakp04]

any clude bout emo portrait code?

[johndpope]

not sure - the contention was around the warping code I actually draft this - warp composition which may be fine https://github.com/johndpope/MegaPortrait-hack/pull/55 the direction from imf is worth looking into - it claims to transcend warping altogether. i get stuck on this one https://github.com/johndpope/IMF/issues/25

while there's interest here - from digging around stable diffusion from what i can tell - tiktok just cherry pick code from magicdance https://github.com/Boese0601/MagicDance - has training code

https://github.com/bytedance/X-Portrait - this quality is quite good. but no training code it maybe possible to retrofit this https://github.com/Boese0601/MagicDance/issues/43

[mantrakp04]

how did you use claude opus to reverse engineer, am curious if we could try improving a model architecture using claude

[johndpope]

get the pro subscription and jam in as much RELEVANT content /context - code spits out ask it a better question better code spits out.